This model is not VAE neutral, and it's baked-in VAE will make washed-out pics (but it works). I recommend downloading and using any of the external avaible VAEs, in my case I use:

Stable Diffusion official VAE (my choice): https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main.
NAI VAE (works fine) - Cannot link to it.

USE RECOMMENDATIONS:

NSFW biased model. But awesome at SFW too. Use it at your own risk of getting unprompted explicit images some times.

This model works better with Natural Language Prompts under 75 tokens. You don't need tag arrays, nor embeddings, detail LORAs or big negatives, in fact small negatives or no-negative works the best. Not all of the images provided have their prompts, but check those that have them to get an idea of what is a good prompt for this model. Also I recommend using a photographer or an artist of your choice (the model has an amazing amount of knowledge about them and you don't need to stick to Greg Rutkowski, there's plenty more out there).

For example you can try: "photo of a young woman with freckles and a red-haired mane, wearing a copper armor at a desolate desert, captured by kristin-lee moolman." or "digital art of a furry calico female, naked amidst a dark spaceship, created by mindmachine." and only the sentence "a render." for the negative (or with no negative at all) - Look at the provided examples with metadata, and build from there.

I have tried to make it as easy to use as possible, however it doesn't react well to many of the "Huge Positive - Huge Negative - Array of contradicting tags" that many people uses... For me it's so simple to use, but it may not be the easiest model depending on your style, give it some time to get used to it :)

The model works well at native 1024x1024 resolution without any High Res. Fix, nor any need for a Face Restoration or anything alike. However it is more consistent if you work in the 768x768 area and then use High Res Fix to reach the 1024x1024 resolution. It also works well for other aspects ratios such as 768x1152 and 1152x768.
I usually use "Euler a" or "DPM++ 2M Karras" at 20 Steps, sometimes "Heun" at 12 Steps. CFG anywhere between the 7-10 area, but any reasonable number will work.
It does not have, nor it needs, any particular trigger words.
CLIP SKIP 1!!! (It is totally untested how does the model perform at Clip Skip 2 or higher, but I assure you that it's not needed for making anime with this model or anything here).

Most of the images with prompts from the preview have been made by @Hoblin, @DarkAesop and me.

Also tested and perfectly compatible with my Ultra High Res. Fix; which will be avaible again soon at our Ko-Fi. There's also a more advanced prompting guide there (although a bit Legacy now): https://ko-fi.com/promptmarinersai

Check any more info or questions at our private Discord here: https://discord.gg/z88HpDwbGq

MODELS (there are two possibilities):

Requiem: You should be using this model. It might not be the easiest one to prompt depending on your style. But you'll get used to it soon and is the most powerful one. It is extremelly good both at photorealistic images, anime style, art or even Furry!
Requiem-Lite: This uses some different weights distribution and it's a bit more forgiving to incoherent prompts and easier to get photorealistic images at it, in exchange it has lost some artistic and anime-like capabilities as well as it has worse artists/photographers knowledge.

VERSIONS (both models have 2 versions):

Standart: Intended for general use.
Inpainting: Intended to use in the inpainting section of the img2img tab, it is mixed with original SD 1.5 inpainting model and is much more coherent whith masking areas. Only select it for that purpose. There's no need to use any additional file.

All of the files are provided in safetensors format. I have never worked in the mix with .ckpt format, nor with unpruned files, so they don't exist. If you need any of them feel free to request them.

ABOUT:

Requiem is my latest attempt of adding value to the SD 1.5 ecosystem, the name comes from the certainty that it will be my last model for this SD version, and I have tried to make my best creation so far.

Despite being "just" a merge, not a trained checkpoint, it is extremelly different in it's making to any of my previous model. I have been using to it's fullest a combination of the newest merging techniques called "Block Merging", "Bayesian Merging" and "Rebasin", and this model is the result of over two months of full dedication, manually scoring the images from the different stages of the merge 12 hours a day, designing tailored strategies to add specific properties from countless models without disrupting the core properties already attained.

As such it has reached a point where it can move from the photoreal area to the digital artwork space just by sheer positive prompting, and stay coherent at it. It is able to make fantasy creatures, monster-girls, landscapes, interiors, action scenes and what-not without the assistance of any LORA, embedding, or any other trick. I have tried it to be capable of Hardcore Sex as well, but I admit that it is still mediocre at that task, and it is a bit biased towards female anatomy, so getting male genitalia is yet a bit tricky, but totally doable without great pain.

As with my other models I have tested most intensively how good it is at making erotic and artistic nudes. But I have tried many other types of prompts too, some of them borrowed.

If assisted with any LORA it can probably go wherever you want. I merged it with most of any free model out there.

As such, it is provided without any license but the original SD one and a Share-Alike restriction. So you can use it for whatever private or commercial purpose you want, merge with it, create and sell any derivate work.

CREDITS

My deepest gratitude to the team from MEH for their tools, support and patience; without their work this would not have been possible. If you are interested in some more technical aspects of merging or may be interested in learning this new methods I totally recommend you visiting their GitHub repos:

https://github.com/s1dlx/sd-webui-bayesian-merger - For bayesian merging.

https://github.com/s1dlx/meh - For everything related to rebasin, block merging, and weight interpolation.

Also some big thanks to all the SD users and creators community. Sadly the list of models used in this blend is long and lost.

Let's all have fun and play together!